Professional Documents
Culture Documents
Abstract— The vulnerabilities of multi-agent-system (MAS) attacks [7]–[9]. Among these attacks, Dos attacks often occur
become a critical issue for cybersecurity. The article investigates in practical multi-agent applications such as formation control.
the formation control problem for MASs under multi-channel They block part or all of the transmission of information
denial-of-service (DoS) attacks. In this article, the attacks on
each channel are independent, while most of the existing results by attacking communication channels, which will result in
show that DoS attacks are the same on all channels. Without loss formation failure [10].
of generality, we consider multi-channel DoS attacks are imposed There are many excellent results in MAS consensus research
on a leader-follower MAS. Firstly, we propose a distributed [11]–[13]. Shen et al. [11] mainly focus on the distributed
formation control protocol to achieve the desired formation in consensus protocol design with considering the transmission
the presence of DoS attacks. A translation-adaptive method is
considered to adjust the interaction weights among neighboring nonlinear problems. Based on neural network and some other
agents online. Furthermore, a performance guarantee is derived schemes, MAS consensus is studied [12]. Adaptive forma-
based on the state information, and hereafter state errors among tion control for MASs has been explored in recent years.
all agents can be regulated. Moreover, we derive the sufficient By means of a model reference control approach, a robust
conditions for system stability w.r.t the controller gain and the adaptive formation controller is used to steer the vehicles
allowable attack duration in the form of linear matrix inequalities
(LMIs). Finally, simulation results are given to illustrate the into a formation pattern [13]. Different from the above, this
effectiveness of the proposed method. article considers that the weight of action between agents
can be adjusted adaptively. The adaptive weight adjustment
Index Terms— Denial-of-service attack, multi-agent system,
adaptive formation, guaranteed-performance, linear matrix lays a foundation for task assignment and formation structure
inequality. optimization. However, these methods are not enough to deal
with the threat of DoS attacks.
I. I NTRODUCTION Guarding the MASs against malicious DoS attacks becomes
a new research topic with related works [14]–[16]. Zhang
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.
3392 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 69, NO. 8, AUGUST 2022
MASs based on the analysis of frequency and duration of DoS attacks. Then, a new translation-adaptive strategy is
DoS attacks. One step further than [21], a distributed secure designed to implement performance constraints. In addition,
control protocol is proposed in [24], and the research object the sufficient conditions of controller gain, an upper bound
is extended to MASs. Heritaging the results in [24], She et of performance function, and attack intensity are converted
al. design a resilient control protocol in [25] to maximize to LMIs. Finally, we consider the various attack models and
attack frequency and duration. Although booming results [21], obtain the corresponding control gain and attack duration
[24], [25] are provided on the security consensus control under different decay rates. Based on the above elaboration,
of MASs, from the perspective of the attacker, all channels we summarize the main contributions of this article as follows:
are considered as one channel. In this article, as the main 1) A guaranteed performance formation control protocol
motivation, we focus on multi-channel DoS attacks which can with adaptive weights under DoS attack is proposed.
occur on any channel at any time. 2) A translation-adaptive strategy is introduced in the con-
For the multi-channel DoS attacks problems, Lu et al. [16], struction of the Lyapunov function to properly integrate
[26], [27] study the resilient control problem of CPSs or MASs performance constraints.
based on the method of state feedback and state estimation. 3) Sufficient conditions for the controller gain, decay rate,
Their results valid for not only periodic DoS attacks, but and attack intensity are given in the form of LMIs.
also when each communication channel is attacked indepen- The rest of the article is organized as follows: Section II
dently. In [16], the research object is extended to MASs, presents the preliminaries and problem formulation. In section
and they design security consensus controllers based on both III, the formation protocol is designed. Section IV gives the
state feedback and state estimation for multi-channel attacks. stability analysis and gets the sufficient condition of DoS
Resiliency is realized based on the method of average state attack duration. To demonstrate the feasibility of the adaptive
consensus. However, the weights between MASs are fixed formation protocol, the simulation results and analysis are
in their works. Different from [16], we consider an adaptive given in section V, and Section VI concludes the paper.
formation for multi-channel DoS attacks, in which the weights Notation: Rn×n denotes the n-dimensional Euclidean space.
of interconnections between agents can be adjusted online. The superscript T stands for matrix transposition, and S > 0,
Multi-channel DoS attacks are independent of each other, and S < 0 denote positive-definiteness and negative-definiteness.
vary in frequency and duration. H e(A) = A + A T , λmax (A) and λmin (A) denote the largest
In addition, the constraint problems should be given more and smallest eigenvalues of A, respectively. Given a vector
consideration for the secure consensus control of practi- v i ∈ Rn , v i is the Euclidean norm of v i . N is the
cal MASs. Several common constraint problems have been set of natural numbers. 0 and I denote a zero matrix and
studied including communication, input saturation, energy unit matrix with appropriate dimensions. The N-dimensional
constraints [28]–[31] and performance constraints [32]–[34]. column vector with all components 1 is represented by 1 N . For
Among the above constraints, the performance constraint as a interval D(t1 , t2 ), |D(t1 , t2 )| is its length over [t1 , t2 ). Given
basic constraint problem is placed in the scope of our works. two sets 1 and 2 , 1 \2 is the relative complement of
When the formation is realized, the tracking error is guar- 2 in 1 . ⊗ and ∗ represent the Kronecker product and the
anteed to be within limits, inspired by [35]. [36] studies the symmetric terms of symmetric matrices, respectively.
unconstrained consensus problems based on a scaling-adaptive
strategy in the absence of attacks. They eliminate the influence II. P RELIMINARIES AND P ROBLEM F ORMULATION
of non-zero eigenvalues of the Laplacian matrix by adding
scaling factors. However, the exact value of the scaling factors A. Graph Theory
are difficult to determine. When we consider performance A leader-follower graph is defined as G(V, E) where V and
constraints, a precise scaling factor must be derived for the E ⊆ V × V are the set of vertices that represents the local
global cost function, which means the utilized in scaling factor agents, and the set of edges that stands for communication
in [36] are no longer suitable. links, respectively. An edge (i, j ) ∈ E denotes that j can
Thus, it is still open to jointly considering the guaranteed obtain information from i .
performance, adaptive formation problem under DoS attacks. E f is the set of edges between follower -follower. The set
The following interesting and challenging problems need to of edges between the leader and follower is denoted by El .
be studied. ωi j (i = j ) represents the action weight of ( j, i ), which is from
agent j to agent i . A MAS with one leader and N followers
1) How to achieve distributed formation under DoS are studied, where the leader is labeled 0, and the followers
attacks, and to guarantee performance among all agents, are labeled 1, 2, 3 · · · , N. The index set Ni = { j : (i, j ) ∈ E}
simultaneously. represents the neighbor set of agent i .
2) How to deal with various attack models come from We define 0 − 1 Laplacian matrix as L = [li j ] ∈
attacks on multiple communication channels. N
R(N+1)×(N+1) with lii = j =0, j =i ωi j,0 , l i j = −ωi j,0 =
3) How to derive the controller gain, attack duration, and
−1 if (v i , v j ) ∈ E and li j = 0, otherwise. A variable Laplace
the upper bound of performance guarantee.
matrix L(t) = [li j (t)] ∈ R(N+1)×(N+1) is defined with lii (t) =
N
In this article, we consider a leader-follower MAS under j =0, j =i l i j ωi j (t) and l i j (t) = −l i j ωi j (t) where ωi j (t) ≥ 1
DoS attacks. Firstly, a guaranteed performance formation is designed later. In addition, the nonzero eigenvalues of L
protocol with adaptively adjusting weights is proposed under and L(t) are positive.
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.
PAN et al.: ADAPTIVE FORMATION FOR MULTIAGENT SYSTEMS SUBJECT TO DENIAL-OF-SERVICE ATTACKS 3393
B. System Description
Given a leader-follower MAS with a leader and N followers,
the dynamic model of the leader is presented as:
ṗ0 (t) = v 0 (t), v̇ 0 (t) = α p p0 (t) + αv v 0 (t),
where p0 ∈ Rn and v 0 ∈ Rn are the positions and velocity
vectors of the leader, respectively. α p and αv are known Fig. 1. Example of DoS signal. On and off mean zero and one, corresponding
to communication allow and disallow.
damping constants. The follower’s dynamic model is given
as:
ṗi (t) = v i (t), v̇ i (t) = α p pi (t) + αv v i (t) + u i (t), Example 1: Three DoS attacks occur within 0 to 15 sec-
onds, that is [s, t] = [0, 15] as shown in Fig. 1. Off-to-On
where i ∈ F = {1, 2, · · · N}. pi ∈ Rn and v i ∈ Rn transitions are represented as ↑. Similarly, ↓ stands for on-
represent the position vector and velocity vector of the i to-off transitions. Off/on transitions occur at t = 1s, 6s, 12.5s
agent, respectively. The formation structure for MAS is speci- and the corresponding time intervals are 3s, 4s, 2.5s, respec-
fied by a vector function f (t) = [ f 0T , f1T , · · · f NT ]T , where tively. This yields: D(0.5, 9) = [1, 4] ∪ [6, 9), D(11, 12) = ∅,
f 0 = 0 ∈ R2n×2n , fi = [ f ipT , fivT ]T ∈ R2n×2n (i ∈ F) D(9, 15) = [9, 10] ∪ [12.5, 15).
is the piecewise continuously differentiable formation vector For DoS attacks, the following Assumption is given.
for follower i with f ip and f iv being the components of Assumption 1: [16] (DoS Duration) There is ξi j > 0 and
f i corresponding to the position and velocity, respectively. 0 < μi j < 1 such that
Defining x 0 = [ p0T , v 0T ]T ∈ R2n , x i = [ piT , v iT ]T ∈ R2n ,
ψ0 = x 0 − f 0 , ψi = [ψipT , ψ T ]T = x − f , it has
iv i i |D(i, j ) (s, t)| ≤ ξi j + μi j (t − s), (4)
ψ˙0 = Aψ0 + A f 0 − f˙0 , where D(i, j ) (s, t) stands for channel (i, j ) suffers the union
of DoS attack over [s, t). μi j represent the attack intensity.
ψ̇i = Aψi + Bu i + A f i − ḟ i , (1)
The set of channels attacked at time t is defined as
0 1 0
where A = ⊗ In ∈ R2n×2n and B = ⊗ In ∈ (t) = {(i, j ) ∈ E|t ∈ D(i, j ) (0, +∞)}. (5)
α p αv 1
R2n×m are system matrices. u i (t) ∈ Rm×2n is the formation The union of two times interval sets,where one is the set of
protocol to be designed. channels subjected to the DoS attack and the other is the set
The control target of adaptive formation of MASs is to of channels not subjected to the DoS attack, is described as
make limt →+∞ x i − fi − x 0 = 0 and Je ≤ Je∗ by designing
appropriate control protocol. Je and Je∗ are the performance
(t1 , t2 ) = (∪(i, j )∈D(i, j ) (t1 , t2 ))∪(∩(i, j )∈/(i, j ) (t1 , t2 )).
index and its upper bound, which will be introduced later. (6)
|E f |
C. Attack Description Remark 1: There are 2 2 +|El |
attack models from no chan-
In this paper, it is assumed that only the transmission nel attacked to all channels attacked. Interval [t1 , t2 ] contains
|E f |
channels are attacked, that is, the control scheme is not able 2 2 +|El | sub-intervals
(t1 , t2 ). Note that
to obtain relevant information from its neighbors. We consider
∪∈E f
(t1 , t2 ) = [t1 , t2 ],
that the communication between followers is undirected and
the communication between leader and followers is directed. D(i, j ) (t1 , t2 ) = ∪∈E f ,(i, j )∈
(t1 , t2 ).
Since the location and time of occurrence are arbitrary, and
|E f |
each channel is likely to be attacked, there are 2 2 +|El | D. Guaranteed Performance Formation Protocol
elements in the set composed of all attack models. We let When DoS attacks the communication channels of the MAS,
{h κ |h κ ≥ 0, κ ∈ N} denote the sequence of DoS off/on the communication is interrupted, and the information in this
transitions. The time instant h κ describes a transition from period is discarded. To ensure the performance of formation,
zero to one, which represents communication connectivity and the performance constraint problem is considered under DoS
communication interruption. Then, attacks. Inspired by [16], [37],a new guaranteed performance
formation protocol is proposed as follows:
Hκ := {h κ } ∪ [h κ , h κ + sκ ), (2)
u i (t) = K u ξi − α fi + f˙iv ,
stands for the κth time-interval, of a length sκ > 0, over which
DoS attack exists. Given t > s > 0, let ξi (t) = ωi0,0 ωi0 (t)(ψ0 − ψi )|(0,i)∈/ (t )
+ (ωi j,0 ωi j (t)(ψ j − ψi ))
D(s, t) = ∪n∈N Hκ ∩ [s, t],
j ∈Ni ,i =0,( j,i)∈
/ (t )
(s, t) = [s, t]\D(s, t). (3)
ω̇i j (t) = ω̇ j i (t) = (ψ j − ψi ))T K ω (ψ j − ψi )),
In other words, D(s, t) and (s, t) are the set of time instants N +∞
where communication denied and allowed for time-interval Je (t) = ωi0 (ψ0 − ψi )T Q(ψ0 − ψi )dt
[s, t], respectively as illustrated in Example 1. i=1 0
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.
3394 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 69, NO. 8, AUGUST 2022
1
N N +∞
one obtain
+ ωi j (ψ j − ψi )T Q(ψ j − ψi )dt,
2 0
i=1 j =1 0 W −1
L(t)W = ,
(7) L f f (t) + f l (t)
0
where i, j = 1, 2, · · · , N, α = [α p , αv ]. Q = Q T ∈ W −1 L (t)W = ff fl .
L (t) + (t)
R2n×2n > 0 is the gain matrix, and Je represents the practical
performance function of the whole MAS. K u ∈ Rm×2n and
K ω = K ωT ∈ Rn×2n are the controller gain matrix and adaptive Because at least there must be a transmission path from
weight gain matrix, which are to be designed. The core idea the leader to every follower, L f f (t) + f l (t) is positive
ff fl
of formation control protocol with guaranteed performance is definite and symmetric. In addition, L (t) + (t) is also
to achieve formation while determining an upper bound Je∗ of a symmetric matrix. There exists an orthonormal matrix
performance index Je . The definition of ωi j,0 and ωi j (t) are W̃ such that W̃ T (L f f (t) + f l (t))W̃ =
f f (t) > 0.
ff fl ff
given as Definition 1. Similarly, ν T (L (t) + (t))ν =
(t) > 0, where
Definition 1: ωi j,0 is the 0 − 1 weight of agent j to agent i . ν = [ν1 , ν2 · · · ν N ]. νi is the orthogonal eigenvector of L
If agent i can receive information from agent j , ωi j,0 = 1(i = corresponding to λi . We define
j ); otherwise, ωi j,0 = 0. ωi j (t) is the adaptive change coef- T
(W −1 ⊗ In )ψ(t) = ψ0 (t) ψ̂(t) ,
ficient, and its initial value is ωi j (0) = ω j i (0) = 1. ω̇i j (t) =
ω̇ j i (t) and ωi j (t) = ω j i (t) are obtained from (7). Due to where
ωi j (t) is bounded, a reasonable assumption ωi j (t) γi j is T
described. According to the definition of 0 − 1 action weight ψ̂(t) = ψ1 (t) − ψ0 (t) · · · ψ N (t) − ψ0 (t) .
ω j i,0 and its adaptive coefficient ω j i (t), the 0-1 Laplace matrix
If (I N+1 ⊗ H1 ) f (t) − (I N+1 ⊗ H2) f˙(t) = 0, there is
L and Laplace matrix L(t) are given as:
˙
ψ̂(t) = (I N ⊗ A − (L f f (t) + f l
0 0
L = , ff fl
− L (t) − ) ⊗ B K u )ψ̂(t).
−L f l L f f + f l (9)
0 0 Theorem 1 is given to get the sufficient conditions of the
L(t) = ,
−L f l (t) L f f (t) + f l (t) formation protocol.
T
L f l = ω10,0 ω20,0 ω30,0 · · · ω N0,0 , Theorem 1: For a connected graph with agents (1), given
T decay rates β, if there are positive symmetric definite X, M
L f l (t) = ω10,0 ω10 (t) · · · · · · ω N0,0 ω N0 (t) such that
⎡ N ⎤ √
i=1 ω1i,0 −ω12,0 · · · −ω1N,0
f f ⊗ XT
⎢ −ω21,0 N ⎥ < 0,
i=1 ω2i,0 · · · −ω2N,0 ⎥
(10)
Lff = ⎢ ⎣ ⎦ ∗ −I N ⊗ M
··· ··· ··· ···
N the inequality is guaranteed
−ω N1,0 −ω N2,0 · · · i=1 ω Ni,0
⎡ N ⎤
i=1 ω1i,0 ω1i (t) · · · −ω1N,0 ω1N (t) V̇ (t) < βV (t). (11)
⎢ .. ⎥
L f f (t) = ⎣ ··· . ··· ⎦ where
N
−ω N1,0 ω N1 (t) · · · i=1 ω Ni,0 ω Ni (t)
N
V (t) = ψ̂ T (t)(I N ⊗ P)ψ̂ (t) + 2γ (γ0i − ω0i (t))
where f l = di ag(L f l ), f l (t) = di ag(L f l (t)), L and L(t)
i=1
are symmetric matrices.
N
N
(ωi j,0 ωi j (t) − ωi j,0 )2
+ ,
III. A DAPTIVE F ORMATION P ROTOCOL D ESIGN 2
i=1 j =1
In this section, we will give the sufficient conditions of and γ ≥ γ0i ≥ ω0i (t), = I N ⊗ H e(AX −β X)−2(
f f −
the formation protocol design scheme. Substituting formation
+ γ I ) ⊗ B R B T , X = P −1 , M = Q −1 , and ∈ E, under
ff
protocol (7) into (1) yields
the distributed formation controller (7) with K u = R B T P
ψ̇(t) = (I N+1 ⊗ A − (L(t) − L (t)) ⊗ B K u )ψ(t) and adaptive weight gain K ω = P B R B T P. The performance
function upper bound satisfies
+ (I N+1 ⊗ H1 ) f (t) − (I N+1 ⊗ H2) f˙(t), (8) +∞
∗ N −1TN
01 10 Je = ψ (0)
T
⊗ Pψ(0) + max(β) V (t)dt
where H1 = , H2 = . L(t) is given in Defini- −1 N I N 0
00 00 +∞
tion 1 and L (t) is defined as L(t) with wi j,0 (( j, i ) ∈
/ (t)) N −1TN
+ 2γ ψ T (t)( ⊗ P B R B T P)ψ(t)dt,
replaced by 0. Furthermore, let 0 −1 N I N
(12)
1 0 where V (t) is given as (A.13).
W = ,
1N IN Proof: See Appendix A for details.
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.
PAN et al.: ADAPTIVE FORMATION FOR MULTIAGENT SYSTEMS SUBJECT TO DENIAL-OF-SERVICE ATTACKS 3395
TABLE I
D ECAY RATES β
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.
3396 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 69, NO. 8, AUGUST 2022
TABLE II
G UARANTEED -P ERFORMANCE C OST
Fig. 5. Trajectories of (ψi ) with applying the scheme of [37]. The states
corresponding to (a), (b) and (c) are ψip X , ψipY and ψip Z . The states
corresponding to (d), (e) and (f) are ψiv X , ψivY and ψiv Z , respectively.
Fig. 6. Position state snapshots of all the agents at different moments for
MASs under DoS attacks.
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.
PAN et al.: ADAPTIVE FORMATION FOR MULTIAGENT SYSTEMS SUBJECT TO DENIAL-OF-SERVICE ATTACKS 3397
Fig. 7. Performance function and guaranteed performance upper bound. Fig. 10. Trajectories of (ψi ) with applying the proposed scheme. The
states corresponding to (a), (b) and (c) are ψip X , ψipY and ψip Z . The states
corresponding to (d), (e) and (f) are ψiv X , ψivY and ψiv Z , respectively.
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.
3398 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 69, NO. 8, AUGUST 2022
Fig. 13. Velocity state of all the agents under DoS attacks. The figures above
and below show the change in velocity in the X and Y directions, respectively.
that is, Je < Je∗ . The performance function and its upper
bound are described in Fig.14. The left vertical axis represents
the actual value of the performance function, and the right
index of the scheme we proposed is less than the scheme
vertical axis represents the upper bound of the performance
of [37].
function. The performance function is far less than its upper
Fig. 10 and Fig. 11 describe the state trajectories ψi of
bound, and the performance guarantee is verified. Fig. 15 is the
the scheme in this paper and the scheme of [37], where
time-varying curve of the adaptive weight change coefficient
the meaning of the description is the same as Fig. 4 and
in the case, which converges to a finite value eventually. Thus,
Fig. 5. In the case of the same attacks and parameters,
the formation can be formed with proposed scheme and it has
the proposed scheme only needs 3s to achieve consensus,
better performance and less time to achieve formation than the
while the scheme of [37] needs close to 30s. The proposed
scheme of [37].
scheme can achieve consensus formation in less time. Fig. 12
and Fig. 13 show position and velocity trajectories of four
agents, where the initial positions of all agents are marked by VI. C ONCLUSION
circles and the final positions are marked by hexagon, pluses, A guaranteed performance adaptive formation protocol is
hexagram, pentagram. Fig. 12 shows the position state change proposed for leader-follower MASs under DoS attacks based
trajectory of all agents. The three followers form an equilateral on time-varying edge weights. Each channel is attacked inde-
triangle and can circle around the leader under DoS attacks. pendently and randomly. Sufficient conditions for the control
All followers move in a circle around the leader with radius protocol and DoS duration are given through LMIs. Two sim-
r = 10. Fig. 13 is the velocity trajectories in the X and Y ulation scenarios are carried out and compared with existing
direction. From the velocity trajectories changes, we can see article to illustrate the effectiveness of the proposed scheme.
that the follower moves in a circle around the leader in a fixed In the future, we will extend the existing results to issues such
formation. The guaranteed performance cost Je∗ = 3581.7, as event-triggered formation.
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.
PAN et al.: ADAPTIVE FORMATION FOR MULTIAGENT SYSTEMS SUBJECT TO DENIAL-OF-SERVICE ATTACKS 3399
A PPENDIX
N
N
(ωi j,0 ωi j (t) − ωi j,0 )2
> 0, (A.7)
A. Proof of Theorem 1 2
i=1 j =1
Introducing translation factor γ > 0, the Lyapunov function we can obtain
is chosen as
ψ̂ T (I N ⊗ P)ψ̂ < V (t),
N
V (t) = ψ̂ (t)(I N ⊗ P)ψ̂ (t) + 2γ
T
(γ0i − ω0i (t)) that is,
i=1
N
N V̇ (t) < βV (t). (A.8)
(ωi j,0 ωi j (t) − ωi j,0 )2
+ , (A.1)
2 Next, consider an guaranteed-performance index such as (7),
i=1 j =1
N +∞
where γ ≥ γ0i ≥ ω0i (t), that is V (t) > 0. It follows from (9)
Je (t) = ωi0 (ψ0 − ψi )T Q(ψ0 − ψi )dt
that 0
i=1
N N
V̇ = ψ̂ (t)(I N ⊗ H e(P A)ψ̂(t) − ψ̂ (t)(L (t) + (t) 1 +∞
T T ff fl
+ ωi j (ψ j − ψi )T Q(ψ j − ψi )dt
ff fl
N
2 0
− L (t)− (t)) ⊗ H e(P B K u ))ψ̂(t)−2γ ω̇0i (t) i=1 j =1
+∞
i=1
= ψ̂ T (t)((L f f + f l ) ⊗ Q)ψ̂(t)dt,
N
N
0
+ (ωi j,0 ωi j (t) − ωi2j,0 )ω̇i j (t). (A.2) h̄
i=1 j =1 Je,h̄ < ψ̂ T (t)((L f f + f l ) ⊗ Q)ψ̂(t)dt
0
According to (7), it has h̄ h̄
= dt − V (h̄) + V (0) + β V (t),
N
N 0 0
(ωi j,0 ωi j (t) − ωi2j,0 )ω̇i j (t) |( j,i)∈/(t )
where = ψ̂ T (t)(L f f
+ f l)
⊗ Q ψ̂(t) + V̇ (t) − βV (t).
i=1 j =1
ff ff
Define (W̃ T ⊗ In )ψ̂(t) = ψ̃(t), that is, ψ̂(t) = (W̃ ⊗ In )ψ̃(t),
= 2ψ̂ T (t)((L f f (t) + f f (t) − L (t) − (t)) and we get
ff fl
− (L f f + f l − L − ) ⊗ K ω ψ̂(t), = ψ̂ T (t)((L f f + f l ) ⊗ X Q X + I N ⊗ (H e(AX)
ff
and it has − β X) − 2(γ I N + L f f + f l − L
fl
N − ) ⊗ B R B T )ψ̂(t).
2γ ω̇0i (t) = 2γ ψ̂ T (t)(I N ⊗ K ω )ψ̂(t). (A.3)
(A.9)
i=1
Let K u = R B T P and K ω = P B R B T P, If ψ̃iT (t) ψ̃i (t) < 0, we obtain (10) and
h̄
V̇ (t) ≤ ψ̂ T (t)(I N ⊗ P)(I N ⊗ H e(AX) − 2(γ I N + L f f Je,h̄ < −V (h̄) + V (0) + β V (t). (A.10)
ff fl 0
+ f l − L − ) ⊗ B R B T )(I N ⊗ P)ψ̂ (t).
Applying the (A.1), we get
(A.4)
N
Subtract (I N ⊗ P)β F X (I N ⊗ P) from both sides of inequal- T
V (0) = ψ̂ (0)(I N ⊗ P)ψ̂(0) + 2γ (γ0i − ω0i (0))
ity (A.4), and we get i=1
V̇ (t) − (I N ⊗ P)β F X (I N ⊗ P)
N
N
(ωi j,0 ωi j (0) − ωi j,0 )2
+ |( j,i)∈/(0) ,
≤ ψ̂ T (t)(I N ⊗ P)(I N ⊗ H e(AX) − β F X − 2(γ I N i=1 j =1
2
ff fl
+ L f f + f l − L − ) ⊗ B R B T )(I N ⊗ P)ψ̂(t).
N
V (h̄) = ψ̂ (h̄)(I N ⊗ P)ψ̂ (h̄) + 2γ
T
(γ0i − ω0i (h̄))
Thus, if the following inequality is guaranteed
i=1
I N ⊗ (H e(AX) − β X) − 2(γ I N + L f f
N
N
(ωi j,0 ωi j (h̄) − ωi j,0 )2
ff fl + . (A.11)
+ f l − L − ) ⊗ B R B T < 0, (A.5) 2
i=1 j =1
it has Then, due to ωi j,0 = ωi j (0) = 1 and limh̄→∞ (ω0i (h̄)) = γ0i ,
V̇ (t) < βψ̂ (I N ⊗ P)ψ̂ .
T
(A.6) we get
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.
3400 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 69, NO. 8, AUGUST 2022
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.
PAN et al.: ADAPTIVE FORMATION FOR MULTIAGENT SYSTEMS SUBJECT TO DENIAL-OF-SERVICE ATTACKS 3401
[10] Y. Yang, Y. Li, D. Yue, Y.-C. Tian, and X. Ding, “Distributed secure [31] H. Huang, C. Yu, and Q. Wu, “Autonomous scale control of multia-
consensus control with event-triggering for multiagent systems under gent formations with only shape constraints,” Int. J. Robust Nonlinear
DoS attacks,” IEEE Trans. Cybern., vol. 51, no. 6, pp. 2916–2928, Control, vol. 23, no. 7, pp. 765–791, May 2013.
Jun. 2021, doi: 10.1109/TCYB.2020.2979342. [32] J. Xi, Y. Yu, G. Liu, and Y. Zhong, “Guaranteed-cost consensus for
[11] Q. Shen, P. Shi, J. Zhu, and L. Zhang, “Adaptive consensus con- singular multi-agent systems with switching topologies,” IEEE Trans.
trol of leader-following systems with transmission nonlinearities,” Int. Circuits Syst. I, Reg. Papers, vol. 61, no. 5, pp. 1531–1542, May 2014.
J. Control, vol. 92, no. 2, pp. 317–328, 2019. [33] X. Zhou, P. Shi, C.-C. Lim, C. Yang, and W. Gui, “Event based guar-
[12] Q. K. Shen, P. Shi, J. W. Zhu, S. Y. Wang, and Y. Shi, “Neural networks- anteed cost consensus for distributed multi-agent systems,” J. Franklin
based distributed adaptive control of nonlinear multiagent systems,” Inst., vol. 352, no. 9, pp. 3546–3563, Sep. 2015.
IEEE Trans. Neural Netw. Learn. Syst., vol. 31, no. 3, pp. 1010–1021, [34] Z. Wang, M. He, Y. Liu, Y. Zhao, and X. Fan, “Guaranteed performance
Mar. 2020. consensus for multi-agent systems with nonlinear dynamics,” in Proc.
[13] N. Xuan-Mung and S. K. Hong, “Robust adaptive formation control of 29th Chin. Control Decis. Conf. (CCDC), May 2017, pp. 268–272.
quadcopters based on a leader–follower approach,” Int. J. Adv. Robot. [35] J. Xi, C. Wang, H. Liu, and L. Wang, “Completely distributed
Syst., vol. 16, no. 4, 2019, Art. no. 1729881419862733. guaranteed-performance consensualization for high-order multiagent
[14] Y. Xu, M. Fang, P. Shi, and Z.-G. Wu, “Event-based secure consensus of systems with switching topologies,” IEEE Trans. Syst., Man, Cybern.,
mutiagent systems against DoS attacks,” IEEE Trans. Cybern., vol. 50, Syst., vol. 49, no. 7, pp. 1338–1348, Jul. 2019.
no. 8, pp. 3468–3476, Aug. 2020. [36] Z. Li, G. Wen, Z. Duan, and W. Ren, “Designing fully distributed
[15] M. Mola, N. Meskin, K. Khorasani, and A. Massoud, “Distributed event- consensus protocols for linear multi-agent systems with directed graphs,”
triggered consensus-based control of DC microgrids in presence of DoS IEEE Trans. Autom. Control, vol. 60, no. 4, pp. 1152–1157, Apr. 2015.
cyber attacks,” IEEE Access, vol. 9, pp. 54009–54021, 2021. [37] Y. Yang, Y. Li, and D. Yue, “Event-trigger-based consensus secure
[16] A.-Y. Lu and G.-H. Yang, “Distributed consensus control for multi-agent control of linear multi-agent systems under DoS attacks over multiple
systems under denial-of-service,” Inf. Sci., vols. 439–440, pp. 95–107, transmission channels,” Sci. China Inf. Sci., vol. 63, no. 5, May 2020,
May 2018. Art. no. 150208.
[17] X.-M. Zhang, Q.-L. Han, X. Ge, and L. Ding, “Resilient control design [38] X. Dong, Y. Zhou, Z. Ren, and Y. Zhong, “Time-varying formation
based on a sampled-data model for a class of networked control systems tracking for second-order multi-agent systems subjected to switching
under denial-of-service attacks,” IEEE Trans. Cybern., vol. 50, no. 8, topologies with application to quadrotor formation flying,” IEEE Trans.
pp. 3616–3626, Aug. 2020. Ind. Electron., vol. 64, no. 6, pp. 5014–5024, Jun. 2017.
[18] D. Zhang and G. Feng, “A new switched system approach to
leader–follower consensus of heterogeneous linear multiagent systems
with DoS attack,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 51, no. 2,
pp. 1258–1266, Feb. 2021, doi: 10.1109/TSMC.2019.2895097. Kunpeng Pan (Graduate Student Member, IEEE)
[19] D. Zhang, L. Liu, and G. Feng, “Consensus of heterogeneous linear received the B.Sc. degree in automation from Dalian
multiagent systems subject to aperiodic sampled-data and DoS attack,” Ocean University, China, in 2017, and the M.Sc.
IEEE Trans. Cybern., vol. 49, no. 4, pp. 1501–1511, Apr. 2019. degree in control theory and application from North-
[20] Y. Yuan and F. Sun, “Data fusion-based resilient control system under east Electric Power University, China, in 2020. He is
DoS attacks: A game theoretic approach,” Int. J. Control, Autom. Syst., currently pursuing the Ph.D. degree in control theory
vol. 13, no. 3, pp. 513–520, Jun. 2015. and applications with the School of Automation,
[21] H. S. Foroush and S. Martinez, “On event-triggered control of linear Northwestern Polytechnical University, China. His
systems under periodic denial-of-service jamming attacks,” in Proc. 51st research interests include switched systems, cyber-
IEEE Conf. Decis. Control (CDC), Dec. 2012, pp. 2551–2556. physical systems, and secure control.
[22] S. Hu, D. Yue, Q.-L. Han, X. Xie, X. Chen, and C. Dou, “Observer-based
event-triggered control for networked linear systems subject to denial-
of-service attacks,” IEEE Trans. Cybern., vol. 50, no. 5, pp. 1952–1964,
May 2020.
[23] S. Hu, D. Yue, Z. Cheng, E. Tian, X. Xie, and X. Chen, “Co-design of Yang Lyu (Member, IEEE) received the Ph.D.
dynamic event-triggered communication scheme and resilient observer- degree in control theory and applications from
based control under aperiodic DoS attacks,” IEEE Trans. Cybern., Northwestern Polytechnical University, Xi’an,
vol. 51, no. 9, pp. 4591–4601, Sep. 2021. China, in 2019. He was a Research Fellow with
[24] Z. Feng and G. Hu, “Distributed secure average consensus for linear Nanyang Technological University, Singapore,
multi-agent systems under DoS attacks,” in Proc. Amer. Control Conf. under the STE-NTU Corporate Laboratory
(ACC), May 2017, pp. 2261–2266. from 2019 to 2021. He is currently an Associate
[25] S. Feng and P. Tesi, “Resilient control under denial-of-service: Robust Professor with Northwestern Polytechnical
design,” Automatica, vol. 79, no. 3, pp. 42–51, 2017. University. His current research interests include
[26] C.-L. Zhang, G.-H. Yang, and A.-Y. Lu, “Resilient observer-based information fusion for robotic systems and resilient
control for cyber-physical systems under denial-of-service attacks,” Inf. robotic systems.
Sci., vol. 545, pp. 102–117, Feb. 2021.
[27] A.-Y. Lu and G.-H. Yang, “Observer-based control for cyber-physical
systems under denial-of-service with a decentralized event-triggered
scheme,” IEEE Trans. Cybern., vol. 50, no. 12, pp. 4886–4895, Quan Pan (Member, IEEE) was born in China in
Dec. 2020. 1961. He received the bachelor’s degree in automatic
[28] Y. She and H. Fang, “Finite-time consensus of multi-agent systems with control from the Huazhong University of Science
communication constraints,” J. Huazhong Univ. Sci. Technol., Natural and Technology, Wuhan, China, in 1982, and the
Sci. Ed., vol. 38, no. 11, pp. 60–63, 2010. master’s and Ph.D. degrees in control science and
[29] T. Yang, Z. Meng, D. V. Dimarogonas, and K. H. Johansson, “Global engineering from Northwestern Polytechnical Uni-
consensus for discrete-time multi-agent systems with input saturation versity (NPU), Xi’an, China, in 1991 and 1997,
constraints,” Automatica, vol. 50, no. 2, pp. 499–506, 2014. respectively. He has been a Professor with the
[30] J. Xi, L. Wang, J. Zheng, and X. Yang, “Energy-constraint formation for School of Automation, NPU, since 1998. His main
multiagent systems with switching interaction topologies,” IEEE Trans. research interests include pattern recognition and
Circuits Syst. I, Reg. Papers, vol. 67, no. 7, pp. 2442–2454, Jul. 2020. information fusion.
Authorized licensed use limited to: Concordia University Library. Downloaded on February 18,2023 at 22:44:42 UTC from IEEE Xplore. Restrictions apply.